This industrial implementation investigates, implements and validates automated methods to object-oriented detection of regions-of-interests (ROI) in hyper-spectral data by applying image processing to multiple feature images.
Usually, hyper-spectral data associated with an image is based on selective measurements from one (or only a small number of) image pixels.
With the help of image processing in the local projection domain of hyper-spectral data (feature images) a selection of ROIs is derived for further processing in the hyper-spectral domain. Subsequently, the ROIs are used to select the most interesting spectra from the acquired hyper-spectral data. The algorithms have to be invariant to influences that are inherent to the hyper-spectral imaging process (e.g. motion blur, environmental conditions like temperature and humidity).
The information content of hyper-spectral data is abstracted by features per pixel (i.e. per spectrum) by means of statistical or chemometric (multivariate) analysis. For the segmentation analysis 40 feature channels of the coffee beans data set served as basis for two different approaches to automated segmentation:
- ROI detection by morphological processing: Since each individual channel is affected by noise due to change in lighting conditions or environmental parameters, the general approach to automated segmentation is based on a three step principle: a) adaptive thresholding to identify fore- and background, b) separation and identification of individual objects (i.e. coffee beans) by means of morphological analysis which can be followed by c) a region growing step to deliver masks for the beans objects.
- Watershed analysis: The first step in automated segmentation based on watershed analysis was the automatic selection of the feature channel having the highest contrast (highest standard deviation in pixel values), which was the feature channel of the 2nd derivative for the coffee beans data. Similar to the first approach, an initial binarization is the basis for the following steps. On this binary mask a distance transform is applied which then delivers local maxima by an iterative search method. The local maxima are used as “seeds” for watershed regions. Combining the watershed regions with the binary mask delivers the desired beans ROIs.